Data-Driven Approach Using Semantics for Recognizing and Classifying TimeML Events in Italian

Proceedings of Recent Advances in Natural Language Processing, pages 533–538, Hissar, Bulgaria, 12-14 September 2011. Data-Driven Approach Using Semantics for Recognizing and Classifying TimeML Events in Italian Tommaso Caselli ILC-CNR, Pisa tommaso.caselli@ilc.cnr.it Borja Navarro-Colorado DLSI - Universidad de Alicante borja@dlsi.ua.es Hector Llorens DLSI - Universidad de Alicante hllorens@dlsi.ua.es Estela Saquete DLSI - Universidad de Alicante stela@dlsi.ua.es Abstract We present a data-driven approach for rec- ognizing and classifying TimeML events in Italian. A high-performance state- of-the-art approach, TIPSem, is adopted and extended with Italian-specific seman- tic features from a lexical resource. The resulting approach has been evaluated over the official TempEval2 Italian test data. The analysis of the results shows a posi- tive impact of the semantic features both for event recognition and classification. Moreover, the presented data-driven ap- proach has been compared with an ex- isting rule-based prototype over the same data set. The results are directly compa- rable and show that the machine learning strategy better deals with the complexity of the tasks. 1 Introduction Recognizing and classifying events is a strategic task in order to improve the performance of many NLP applications such as automatic summariza- tion and question answering (Q.A.). In NLP, dif- ferent definitions of event can be found regarding the target application. Recently, TimeML (Puste- jovsky et al., 2003a) introduced a rich specifica- tion language for annotating and classifying events and it has been applied to English documents (the TimeBank corpus (Pustejovsky et al., 2003b)). The SemEval TempEval-1 and TempEval-2 in- ternational evaluation excercises (Verhagen et al., 2007; Verhagen et al., 2010), have provided the NLP community with gold standard resources for comparative evaluations of different systems. In addition to this, TempEval-2 made available TimeML annotated data in languages other than English, namely Italian, French, Spanish, Chinese and Korean. Unfortunately, there were only par- ticipants for English and Spanish. This paper focuses on the recognition and clas- sification of TimeML events in Italian by means of a state of the art data driven approach, TIPSem (Llorens et al., 2010), which obtained competi- tive results in the TempEval tasks for English and Spanish. To the best of our knowledge this is the first data-driven approach which is developed for this language and differs from state-of-the-art approches developed for English for the use of specific lexical semantic features. In particular, the Italian-specific semantic features have been obtained from a semi-automatically built event lexicon which has been derived from the SIM- PLE/CLIPS lexicon (Ruimy et al., 2003), follow- ing the proposal in Caselli (2009). The objec- tives of this paper are (i) evaluating TIPSemIT over the official TempEval-2 data for Italian and assessing the impact of the semantic resource, and (ii) comparing the performance of data-driven to rule-based approaches in Italian over the same data. Section 2 reports a short background on the TimeML specifications. Section 3 describes the adaptation of TIPSem approach to Italian. Sec- tion 4 is devoted to the different evaluation exper- iments introduced above and, finally, section 5 fo- cuses on conclusions and future …

Content truncated to 500 words